Anthropic Claims Unreleased Claude AI Raised Key Riemann Hypothesis Bound to 67.2%

There’s a problem in mathematics that has defeated every human mind that’s ever looked at it seriously. The Riemann hypothesis has sat unsolved since Bernhard Riemann first posed it in 1859, and it’s considered so important that the Clay Mathematics Institute lists it as one of the Millennium Prize Problems, with a prize of around £790,000 (originally $1 million) for whoever finally cracks it. This week, Anthropic says one of its AI models made the largest single-step advance on a related measure on record — without solving the problem itself.

What Anthropic Actually Claims

The company posted to X that an unreleased research version of Claude was set to work on the Riemann hypothesis. It did not solve it. But what it did do, Anthropic says, is push forward a specific lower bound: the fraction of nontrivial zeros of the Riemann zeta function that are confirmed to lie on the so-called critical line.

That bound has sat at 41.6% for some time. Claude pushed it to 67.2%.

Anthropic’s research page describes this as the largest single-step improvement on record for this particular measure. The work was carried out over two sessions inside Claude Code — Anthropic’s coding and reasoning environment — and used a total of 31 million output tokens across those sessions. A paper and formalised proof artefacts have been made available publicly, though the model that produced them has not been released.

Why the Riemann Hypothesis Matters

So what is all this actually about? The Riemann hypothesis concerns the location of what are called the nontrivial zeros of the Riemann zeta function — points where the function equals zero. The hypothesis predicts that all of those zeros lie on a specific vertical line in the complex plane, known as the critical line. If true, it would have profound consequences for our understanding of prime numbers and much of modern number theory.

Nobody has proved it. Nobody has disproved it.

What mathematicians have done over the decades is establish lower bounds — proofs that at least a certain percentage of those zeros must lie on the critical line, even if we can’t yet prove all of them do. The 41.6% figure was the established record. Claude, according to Anthropic, has now moved that needle to 67.2% in what the company describes as a meaningful advance in pure mathematical research.

How the AI Did It

Anthropic frames this as evidence that AI can contribute genuine research-level progress in mathematics, not just assist with calculation or code. The use of Claude Code across two sessions, generating 31 million output tokens, suggests this wasn’t a quick query-and-answer exercise. It was, by the numbers, a hefty and extended piece of computational work.

But the model itself is unreleased. That’s a real limitation when it comes to independent verification. Mathematicians and commentators have been broadly positive about the result, while also being careful to point out that this is not a proof of the Riemann hypothesis. The gap between confirming 67.2% of zeros lie on the critical line and proving that 100% do is not just a numerical gap — it’s a conceptual one of enormous depth.

Some researchers have noted that because the model isn’t publicly available, reproducibility remains limited for now. The paper and proof artefacts help, but the full picture depends on what the model actually did and how.

The Broader Picture for AI and Mathematics

Anthropic’s position is clear: this is about demonstrating that AI can make real progress on hard, unsolved problems in pure mathematics, even when a complete solution is out of reach. And there’s something worth sitting with in that framing. Progress on the Riemann hypothesis, even incremental progress, has historically come in very small steps over very long periods.

Moving a bound from 41.6% to 67.2% in two sessions is not a small step by historical standards.

Yet mathematicians are right to hold the line on what this means. The Riemann hypothesis is not closer to being proved in the sense that most people would understand. What’s changed is one specific measure of our partial knowledge about where its zeros sit. That’s a real result. It’s just not the result the headlines might suggest.

The announcement lands at a moment when AI companies are increasingly pointing to scientific and mathematical research as proof of their models’ reasoning abilities. Whether this result holds up to full peer scrutiny, and what it says about the capabilities of the unreleased model, are questions that the mathematics community will be working through in the weeks ahead.

What This Means for Kent Residents

For most people in Kent, this story is one to watch as a broader indicator of where AI research is heading, rather than something with an immediate local impact. Universities in the region with mathematics or computer science departments — including those engaged in AI research — may find this kind of result relevant to ongoing academic discussions about what AI systems can and cannot do. More broadly, as AI companies like Anthropic push into scientific research, the tools and models they develop will eventually filter into everyday software, services, and industries that Kent workers and consumers already use.